Monocular Underwater Depth Estimation via Style Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Underwater image depth estimation and color correction are challenging due to poor visibility and geometric distortion, with scarce labeled datasets and the need for supervision information, limiting the effectiveness of existing methods in the underwater environment.
Innovation Solution
A monocular underwater image depth estimation and color correction framework based on a deep neural network, comprising a style transfer subnetwork using Generative Adversarial Networks to generate synthetic labeled data and a task subnetwork for collaborative learning, which combines depth estimation and color correction tasks to improve accuracy and adapt to the underwater domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep convolution neural network is used for depth estimation from single color image, then the limitation of imaging related in traditional methods is solved, but the scarcity of labeled underwater image datasets makes training difficult
Solution Approach 1:
The patent uses style transfer network to generate synthetic underwater images with corresponding depth maps by copying and transforming land image datasets into underwater scenarios. This creates artificial labeled data without requiring actual underwater depth measurements, solving the data scarcity problem while maintaining the deep learning approach's versatility.
Solution Approach 2:
The patent performs preliminary style transfer training to generate synthetic underwater datasets before conducting the actual depth estimation task. This preliminary data preparation enables the subsequent depth estimation network to be trained on sufficient labeled data, addressing the quantity issue before the main task execution.
2Measurement precision
If stereo matching technology or depth sensing devices are used, then depth information can be obtained, but the imaging limitations and optical distortion in underwater environment make results unsatisfactory
Solution Approach 1:
The patent replaces physical depth sensing devices and stereo matching mechanical systems with a computational deep learning approach. By using neural networks trained on synthetic data, the system achieves depth estimation without being affected by underwater optical distortion, eliminating the harmful factors that plague traditional imaging-based methods.
3Measurement precision
If existing depth estimation methods are used, then depth information can be inferred, but the lack of effective supervision information limits the ability to build more accurate depth estimation networks
Solution Approach 1:
The style transfer network generates synthetic underwater images with corresponding depth maps, creating artificial supervision information that would otherwise be unavailable in real underwater scenarios. This copied and transformed data provides the necessary labels for training accurate depth estimation networks without requiring actual measured depth data from underwater environments.
Data Source
AI summary
The invention discloses a depth estimation and color correction method for monocular underwater images based on deep neural network, which belongs to the field of image processing and computer vision. The framework consists of two parts: style transfer subnetwork and task subnetwork. The style transfer subnetwork is constructed based on generative adversarial network, which is used to transfer the apparent information of underwater images to land images and obtain abundant and effective synthetic labeled data, while the task subnetwork combines the underwater depth estimation and color correction tasks with the stack network structure, carries out collaborative learning to improve their respective accuracies, and reduces the gap between the synthetic underwater image and the real underwater image through the domain adaptation strategy, so as to improve the network's ability to process the real underwater image.